[Dangers of sub-group analysis and of multiple tests]

Subgroup analyses, "retrospective" statistics, and multiple tests, motivated by the search for explanatory covariables, are frequently used in the medical literature. However, these data manipulations induce modifications of the groups defined by initial randomization, an unverifiable increase of the alpha risk, a reduction of the power of the study and a risk of circular reasoning. Taking the example of an article concerning the hormonal treatment of prostate cancer, several situations are discussed, according to whether or not the study is prospective, whether or not the covariables are defined from the outset, and whether or not the results for the primary endpoint are significant.